model-writtenrun 79 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Yuelin Wang reported that a typanalyze callback could accept an expression during initial lookup but reject it when given the actual statistics target, allowing a NULL pointer to reach extended-statistics builders and crash ANALYZE #1.
- Tomas Vondra initially failed to reproduce the crash with his callback #2, then corrected it to initialize valid statistics through std_typanalyze() and supplied a working reproducer, while noting that such callback behavior was unlikely in practice #3.
- Shihao Zhong turned Tomas's reproducer into a fix for failed expression analysis and regression tests, supplying development and older-branch variants #4.
- Michael Paquier identified the second unchecked caller, build_expr_data(), questioned duplicated warnings, and rejected a redundant single-expression test #5; Shihao confirmed the second crash path and added its protection while consolidating warnings and removing the duplicate test #6.
- Michael requested a dedicated test for the second failure path #7, which Shihao supplied with a callback that rejects the second build-time call #8; subsequent review led to shorter comments and an inline warning helper #9 #10.
- Michael committed #10 as 4f89416c9d with light cleanup, renamed callbacks, test-module documentation, and stronger tests checking that no statistics-data row exists.
- Michael's final report confirmed that the fix had landed and been backpatched through PostgreSQL 14, whose test retained the stxdexpr check; no further work remained pending in the thread #11.